Kill Or Die: Moral Judgment Alters Linguistic Coding of Causality. Journal of Experimental Psychology: Learning, Memory, and Cognition

Home , Blurry

Journal of Experimental Psychology: Learning, Memory, and Cognition Kill or Die: Moral Judgment Alters Linguistic Coding of Causality Julian De Freitas, Peter DeScioli, Jason Nemirow, Maxim Massenkoff, and Steven Pinker Online First Publication, February 2, 2017. http://dx.doi.org/10.1037/xlm0000369

CITATION De Freitas, J., DeScioli, P., Nemirow, J., Massenkoff, M., & Pinker, S. (2017, February 2). Kill or Die: Moral Judgment Alters Linguistic Coding of Causality. Journal of Experimental Psychology: Learning, Memory, and Cognition. Advance online publication. http://dx.doi.org/10.1037/xlm0000369 Journal of Experimental Psychology: © 2017 American Psychological Association Learning, Memory, and Cognition 0278-7393/17/$12.00 http://dx.doi.org/10.1037/xlm0000369 2017, Vol. 10, No. 999, 000

Kill or Die: Moral Judgment Alters Linguistic Coding of Causality

Julian De Freitas Peter DeScioli Harvard University Stony Brook University

Jason Nemirow Maxim Massenkoff Harvard University University of California, Berkeley

Steven Pinker Harvard University

What is the relationship between the language people use to describe an event and their moral judgments? We test the hypothesis that moral judgment and causative verbs rely on the same underlying mental model of people’s actions. Experiment 1a finds that participants choose different verbs to describe the major variants of a moral dilemma, the trolley problem, mirroring differences in their wrongness judgments: they described direct harm with a single causative verb (Adam killed the man), and indirect harm with an intransitive verb in a periphrastic construction (Adam caused the man to die). Experiments 1b and 2 separate physical causality from moral valuation by varying whether the victim is a person or animal and whether the harmful action rescues people or inanimate objects. The results show that people’s moral judgments lead them to portray a causal event as either more or less direct and intended, which in turn shapes their verb choices. Experiment 3 finds the same basic asymmetry in verb usage in a production task in which participants freely described what happened.

Keywords: moral cognition, moral psychology, causative verbs, trolley problem, argument structure

Supplemental materials: http://dx.doi.org/10.1037/xlm0000369.supp

In many languages, speakers use distinct verbs for killing and struction, and also do not fit into either moral category. If Alice dying (Haspelmath, 1993). The English verb kill, for example, is caught the lobster and sold it to a cook, people might disagree transitive, requiring a subject and object, as in Alice killed the about whether she killed it or caused it to die. Similarly, if a lobster, whereas the verb die is intransitive, requiring only a general ordered a drone strike that also hit civilians as collateral subject, as in The lobster died. These verbs are not interchange- damage, people could disagree about whether the general killed the ء ء able: Speakers do not say Alice died the lobster or The lobster civilians or caused them to die. In both cases, observers might also killed (to mean that the lobster died). In contrast, there are many disagree about the agents’ moral culpability. other verbs that appear in both forms, such as boil in Alice boiled Here we investigate whether the same considerations that shape the lobster and The lobster boiled (Pinker, 1989, 2007; Gergely & people’s choice of verb, such as killing or dying, also shape their Bever, 1986; Goldberg, 2001; Levin, 1993; Spellman & Mandel, moral judgments. We build on a theory that moral judgment and 1999). verbs rely on the same underlying mental model of people’s Killing and dying are also very different categories in our moral actions (Pinker, 2007). Prior research on verbs found that partic- judgments. Someone could blame Alice for killing a lobster but ipants tend to use a single causative verb (a lexical causative)

This document is copyrighted by the American Psychological Association or one of its alliedshe publishers. is not necessarily to blame if the lobster died. Many real-life when an actor affects an object intentionally and directly, that is,

This article is intended solely for the personal use ofsituations the individual user and is not to be disseminated broadly. are not obligatorily expressed by one or the other con- without an intermediate link in the causal chain consisting of a second actor or of a natural event with causal potency such as electricity, a vehicle, or the weather (Wolff, 2003). In contrast, participants use an intransitive verb as a complement to a generic Julian De Freitas, Department of Psychology, Harvard University; Peter causal verb such as cause or make (a periphrastic construction) DeScioli, Department of Political Science, Stony Brook University; Jason when an actor affects an object via intervening causes. Thus a boy Nemirow, Department of Psychology, Harvard University; Maxim Mas- pops a balloon when he pricks it, but causes the balloon to pop senkoff, Department of Economics, University of California, Berkeley; when he allows it to graze a hot light bulb on the ceiling. This Steven Pinker, Department of Psychology, Harvard University. difference between verb constructions can be explained by an We thank Ben Carter and Bowen Cho for research assistance. Correspondence concerning this article should be addressed to Julian underlying mental model of force dynamics in which people imag- De Freitas, Department of Psychology, Harvard University, William ine that an antagonist exerts force on an agonist through physical James Hall 964, 33 Kirkland Street, Cambridge, MA 02138. E-mail: contact (Pinker, 2007; Wolff, 2007). Foreseeabilty and intention [email protected] matter as well (De Freitas & Johnson, 2015; Malle, 2006): An

1 2 DE FREITAS ET AL.

actor who deliberately lets a balloon graze a hot bulb with the The pure causality and moral–physical conflation hypotheses intention that the balloon burst could be said to pop the balloon. diverge when the physical causal structure of a scenario is kept This previous work examined morally neutral events such as constant while the moral structure is changed. In such instances, moving a marble, extinguishing a candle, and turning on a TV. In the pure causality hypothesis predicts no change in verb choice, this article, we examine the connection between language and because the verbs will reflect only the physical structure. In con- morality by using moral dilemmas as stimuli to simultaneously trast, the moral–physical conflation hypothesis predicts that verb observe verb choice and moral judgment. Specifically, we study choice will mirror a change in moral structure, such that a more two common variants of the trolley problem: the footbridge case in condemnable action is more likely to be described by a single caus- which an actor pushes one person in front of a trolley to save five ative verb. If so, participants will be more likely to choose the people, and the switch case in which an actor flips a switch to causative verb in the footbridge than the switch dilemma; yet, this redirect the trolley toward one person in order to save five people difference would not necessarily occur in physically equivalent sce- (Foot, 1978). Despite the similar tradeoff, most participants judge narios that do not differ in moral wrongness. pushing to be morally wrong but flipping the switch to be morally In Experiment 1b through Experiment 2, we separate physical permissible (Hauser et al., 2007). causality from morality by using the same description of the Researchers have proposed several theories to account for the physical forces that drive an event while changing the level of footbridge-switch difference, and these theories continue to be moral wrongness. We do so by varying whether the victim is a debated (DeScioli & Kurzban, 2009, 2013; Greene, Sommerville, person or an animal and whether the actor rescued five people or Nystrom, Darley, & Cohen, 2001; Kurzban et al., 2012; Mikhail, five inanimate objects. In Experiment 3, we test whether the basic 2007). We suggest that the same force-dynamic mental model that footbridge-switch difference in verb choice also translates to nat- governs the choice of causative constructions might contribute to ural language production by asking participants to explain in the moral difference (Pinker, 2007). In the footbridge dilemma, the open-ended responses their moral judgments for the footbridge and actor directly contacts the victim, fitting the prototype for a causal switch dilemmas. action. In the switch dilemma, the actor only indirectly affects the victim by flipping a switch that redirects the trolley, deviating from Experiment 1a the prototype. Experiment 1a examines the connection between moral judg- If force dynamics underlie the footbridge-switch difference, ment and verbs by observing whether participants choose different then participants will choose different verbs to describe these verb constructions to describe the footbridge and switch dilemmas. morally contrasting scenarios. In Experiment 1a, we measure The pure causality hypothesis predicts that people will show participants’ preference for the transitive causative verb in killed greater use of the causative verb kill in the footbridge dilemma the man over the intransitive verb in caused the man to die in the than in the switch dilemma, mirroring the contrast in their moral footbridge and switch scenarios. A pure causality hypothesis pre- judgments. dicts greater use of the lexical causative kill for the footbridge than In this and subsequent experiments, the main focus is partici- the switch scenario, mirroring the usual difference in judgments of pants’ verb choice. Hence, participants first answered a verb wrongness. This result would point to a common mental model of choice question before providing their moral judgments. In the action underlying two seemingly different phenomena, verb choice present context, participants’ moral judgments are essentially a and moral judgment. manipulation check to confirm that they show the same footbridge- Experiments 1b through 2 consider a second mechanism that switch moral asymmetry that has been replicated many times could align moral judgments with verb choices: in addition to a before (Greene, 2014) or, in Experiments 1b through 2b, that this mental model of causation driving moral judgment, the reverse moral asymmetry has been muted or amplified as intended by the might occur. That is, people who condemn an action might portray modified scenarios. its causal dynamics as more direct and intentional in order to The primary measure of verb choice is participants’ forced- support their accusation, using causative verbs to convey this choice selection between two sentences to describe the actor’s condemnation; we refer to this as the moral–physical conflation behavior: a sentence with the lexical causative kill or a sentence hypothesis. For example, people who want to condemn a drone with a periphrastic causative consisting of a generic causative This document is copyrighted by the American Psychological Association or one of its allied publishers. attack might choose to say the general killed the civilians, whereas added to the intransitive verb die. Second, we ask participants to This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. defenders of the attack might prefer the general caused the civil- rate the accuracy of each sentence separately. Here we are mainly ians to die. In the trolley problem, greater condemnation of the interested in ratings of the lexical causative kill, because previous actor in the footbridge version could motivate greater use of the research found that lexical causatives are more particular to un- causative verb kill. In support of this two-way influence, previous mediated and/or intentional actions, whereas periphrastic caus- research shows that although moral judgments are influenced by atives show less specificity and can be applied to mediated or more basic assessments of causality, intentions, and harm, the unmediated causes (Wolff, 2003). Last, we follow previous re- reverse can also occur: People’s moral condemnation can lead search by also measuring participants’ perceptions of the number them to portray causality, intentions, and harm to support their of events that occurred in the dilemma. moral judgment (Alicke, 1992; DeScioli et al., 2011; Haidt, 2001; Knobe, 2005). For instance, people commonly judge that negative Method side effects are more intentional than positive side effects (Knobe, 2005). However, it is unknown whether these effects extend to We recruited participants using Amazon’s Mechanical Turk, an language, shaping the verbs people choose. online crowdsourcing platform (see, Berinsky, Huber, & Lenz, 2012; KILL OR DIE 3

Buhrmester, Kwang, & Gosling, 2011; Goodman, Cryder, & Cheema, into, C. A woman screamed for help, D. All of the above, E. None 2012; Ipeirotis, 2010; Paolacci, Chandler, & Ipeirotis, 2010). We of the above), and completed demographic items. chose a sample size that provides sufficient power to detect a large effect size, and excluded 16 participants for incomplete responses, Results and Discussion previous participation in similar studies, or incorrectly answering a comprehension question at the end of the study (n ϭ 2, see the Ratings of verb accuracy. Participants rated the causative following text for question wording), yielding a sample of 104 par- verb kill as more accurate in the footbridge condition (M ϭ 5.72, ϭ ϭ ϭ ϭ ticipants (Mage 33 years, 43% female). SD 1.78) than in the switch condition (M 3.72, SD 2.11), Participants were assigned to either the switch or footbridge t(102) ϭ 5.20, p Ͻ .001, d ϭ 1.03. Participants also rated the condition. in the switch condition, participants read the following: statement with the intransitive verb die as more accurate in the footbridge condition (M ϭ 6.24, SD ϭ 1.48) than in the switch One day Adam was walking near some trolley tracks. Suddenly, a condition (M ϭ 5.54, SD ϭ 1.72), t(102) ϭ 2.22, p ϭ .028, d ϭ trolley was quickly approaching out of control. Adam saw that five 0.44. people were standing on a crosswalk in the trolley’s path. Adam was Forced-choice of verb accuracy. In the footbridge dilemma, standing next to a switch that could divert the trolley onto a sidetrack, participants chose the sentence with the causative verb kill 46% of the but there was one man standing on the sidetrack. The five people were time, the intransitive verb die 48%, and neither 6%, whereas in the too far away to hear Adam’s warnings and there was no other way to help them. Adam decided to pull the switch to divert the trolley. As a switch dilemma they chose kill 9% of the time, die 61% of the time, result, the five people escaped unharmed but the one man who was and neither 30% of the time. Chi-square tests showed that participants standing on the sidetrack was killed by the trolley. chose kill significantly more often in the footbridge than in the switch condition, ␹2(1, N ϭ 104) ϭ 17.81, p Ͻ .001, ␸ϭ.41 (see Figure 1). In the footbridge condition, participants read a similar vignette Number of events. There was no significant difference in in which Sentences 4 onward were replaced with the following: whether participants thought multiple events (rather than a single event) had occurred in the footbridge (78%) and switch (83%) . . . Adam was standing on a footbridge over the tracks next to a man conditions, ␹2(1, N ϭ 104) ϭ 0.48, p ϭ .491, ␸ϭ.07. wearing a large, heavy backpack. Adam realized that he could slow Moral judgment. Participants’ moral judgments replicated the trolley and save the five people if he pushed a heavy object in the path of the trolley. The only object that was heavy enough was the previous work as expected. For ratings of moral wrongness, par- man with the backpack. The five people were too far away to hear ticipants rated the actor’s behavior as more wrong in the footbridge Adam’s warnings and there was no other way to help them. Adam condition (M ϭ 4.88, SD ϭ 1.61) than in the switch condition decided to push the man onto the tracks in front of the trolley. As a (M ϭ 2.94, SD ϭ 1.72), t(102) ϭ 5.91, p Ͻ .001, d ϭ 1.17. For result, the five people escaped unharmed but the one man who was the forced-choice question about whether the actor’s behavior was pushed onto the tracks was killed by the trolley. morally wrong (yes or no), participants judged that the actor’s behavior was morally wrong more often in the footbridge condi- Participants rated the accuracy of the following two statements tion (68%) than in the switch condition (19%), ␹2(1, N ϭ 104) ϭ about the scenario on a 7-point scale (scale anchors: very poor 26.04, p Ͻ .001, ␸ϭ.50. description to very accurate description), beginning with a state- In sum, we find that participants’ verb choices mirrored their ment with a causative verb (Adam killed the man) followed by a moral judgments: They were more likely to choose the causative statement with an intransitive verb in a periphrastic construction verb kill in the footbridge scenario than the switch scenario (see (Adam caused the man to die). Participants also answered a forced- Figure 1). choice item about which of these statements best described what happened (Choose the sentence that best describes what happened in this scenario; the answer options were Adam killed the man, Experiment 1b Adam caused the man to die, or neither). Following previous Experiment 1b distinguishes between the possible mechanisms research on causal chains and verbs (Wolff, 2003), we also asked discussed in the introduction that could align verbs with moral participants whether they thought that one or multiple events judgments in the footbridge and switch dilemmas. According to occurred in the scenario, using the same wording from previous the pure causality hypothesis, speakers categorize an event by This document is copyrighted by the American Psychological Association or one of its allied publishers. work (How many events do you think occurred in this scenario? consulting an intuitive model of the physical force dynamics This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. Answer this question casually as if someone asked you how many involved, and this categorization drives both their verb choice and events occurred; the answer options were one event or multiple their moral judgment. According to a moral–physical conflation events.) hypothesis, people have a mental category that combines physical To measure moral wrongness, participants then answered a causation and moral responsibility (an agent foresees, desires, and forced-choice question about whether the actor’s behavior was directly brings about an outcome), so that a moral judgment can morally wrong (yes or no). They also rated how morally wrong the alter their verbal description of an event, holding constant their behavior was on a 7-point scale (scale anchors: not at all morally understanding of the physical dynamics. wrong to extremely morally wrong). Participants then explained in We separate causality from morality by changing the victim in a textbox how they made their decisions for the entire study the trolley dilemmas from a person to a cow. This keeps physical (Please describe how you made your decisions in this study), causality the same as in Experiment 1a, while making the actor’s answered a comprehension question about the scenario, Which, if behavior less morally wrong (at least according to the prevailing any, of the following events occurred in the scenario that you moral standard in the West in which the life of an animal is not read? (answer options: A. A TV was stolen, B. A car was broken morally sacrosanct in the way that the life of a human is). The pure 4 DE FREITAS ET AL.

Figure 1. Participants’ moral judgments and choice of the verb kill (vs. die or neither) for Experiments 1a (a), 1b (b), 2a (c), and 2b (d).

causality hypothesis predicts that verb choice will show the same intransitive verb die did not differ between the footbridge condi- footbridge-switch difference as in Experiment 1a because the tion (M ϭ 6.25, SD ϭ 1.19) and switch condition (M ϭ 5.96, SD ϭ difference in causal structure is the same. Alternatively, the moral– 1.61), t(100) ϭ 1.03, p ϭ .307, d ϭ 0.21. physical conflation hypothesis predicts that participants will show Forced-choice of verb accuracy. In the footbridge dilemma, no footbridge-switch difference in verb choice if there is no participants chose the sentence with the causative verb kill 11% of difference in moral wrongness. the time, the intransitive verb die 79% of the time, and neither 9% of the time, whereas in the switch dilemma they chose kill 4% of Method the time, die 78% of the time, and neither 18% of the time; the proportion choosing kill did not differ between conditions (p ϭ We recruited participants using Amazon’s Mechanical Turk, .272, Fisher’s exact test; see Figure 1).

This document is copyrighted by the American Psychological Association or one of its alliedchoosing publishers. a sample size that provides sufficient power to detect a Number of events. There was no significant difference in

This article is intended solely for the personal use oflarge the individual user and is not to beeffect disseminated broadly. size. We excluded 18 participants for incomplete whether participants thought multiple events (rather than a single responses, previous participation in similar studies, or incorrectly event) had occurred in the footbridge (68%) and switch (71%) answering the same comprehension question as in Experiment 1a conditions, ␹2(1, N ϭ 102) ϭ 0.08, p ϭ .779, ␸ϭ.03. (n ϭ 4), yielding a sample of 102 participants (M ϭ 34 years, age Moral judgment. Participants’ wrongness judgments con- 53% female). The experimental design and stimuli were the same firmed that changing the sacrificed victim to a cow muted the as in Experiment 1a, except the victim was changed from a man to moral differences between footbridge and switch scenarios. For a cow. ratings of moral wrongness, participants rated the actor’s behavior as more wrong in the footbridge condition (M ϭ 1.87, SD ϭ 1.13) Results and Discussion than in the switch condition (M ϭ 1.39, SD ϭ 0.53), t(100) ϭ 2.71, Ratings of verb accuracy. Participants rated the causative p ϭ .008, d ϭ 0.54, although these ratings were much lower than in verb kill as more accurate in the footbridge condition (M ϭ 4.98, Experiment 1a (footbridge: 4.88, switch: 2.94). For the forced-choice SD ϭ 1.95) than in the switch condition (M ϭ 3.94, SD ϭ 2.32), question about whether the actor’s behavior was morally wrong (yes t(100) ϭ 2.46, p ϭ .015, d ϭ 0.49. Participants’ ratings for the or no), wrongness judgments did not differ between the footbridge KILL OR DIE 5

condition (8%) and switch condition (0%; p ϭ .119, Fisher’s exact a result, the five bicycles escaped undamaged but Lucy, who was test). These percentages were much lower than in Experiment 1a standing on the sidetrack, was killed by the trolley. (footbridge: 68%, switch: 19%), confirming that most participants viewed sacrificing a cow to save five people as morally permissible. In the footbridge condition, participants read the following: In short, we found that simply changing the victim from “man” to One day Adam was walking near some trolley tracks. Suddenly, a trolley “cow” was sufficient to alter which verbs participants chose to de- was quickly approaching out of control. Adam saw that five bicycles scribe the physical event (see Figure 1). Specifically, a large majority were standing on a crosswalk in the trolley’s path. Adam was standing on of participants in both scenarios preferred the intransitive verb die a footbridge over the tracks next to a large cow, named Lucy. Lucy was over the causative verb kill. This shows that verb choice is sensitive dearly loved by the community and was well-known for her gentle not only to the causal structure of a scenario but also to the intuitive demeanor and for raising many of the young cows in the town. Adam wrongness of the action, supporting the moral–physical conflation realized that he could slow the trolley and save the five bicycles if he hypothesis, in which people who condemn an action portray its causal pushed a heavy object in the path of the trolley. The only object that was heavy enough was Lucy the cow. The five bicycles were too far away and dynamics as more direct and intentional to support their accusation, there was no other way to save them. Adam decided to push Lucy onto using causative verbs to convey this condemnation. Recall that the the tracks in front of the trolley. As a result, the five bicycles escaped physical difference between the footbridge and switch conditions— undamaged and Lucy, who was pushed onto the tracks, was killed by the pushing the victim or pulling a switch—was identical to the original trolley. scenarios used in Experiment 1a. Yet participants did not tend to choose the causative verb kill for pushing the cow to save five people, Results and Discussion whereas they did for pushing the man to save five people, consistent with the moral–physical hypothesis that diminishing wrongness also Ratings of verb accuracy. Participants rated the causative diminishes the choice of the causative verb. verb kill as more accurate in the footbridge condition (M ϭ 5.77, SD ϭ 1.78) than in the switch condition (M ϭ 4.83, SD ϭ 2.14), ϭ ϭ ϭ Experiment 2a t(145) 2.90, p .004, d 0.48. Participants also rated the intransitive verb die as more accurate in the footbridge condition Experiment 2a further tests the moral–physical conflation hy- (M ϭ 6.52, SD ϭ 0.95) than in the switch condition (M ϭ 5.82, pothesis by changing the circumstances of the dilemmas to make SD ϭ 1.68), t(145) ϭ 3.13, p ϭ .002, d ϭ 0.52. sacrificing the cow morally wrong. Specifically, we changed the Forced-choice of verb accuracy. In the footbridge dilemma, rescuees from people to bicycles, and we personalized the cow. participants chose the sentence with the causative verb kill 27% of We expected participants to judge sacrificing a special cow to save the time, the intransitive verb die 72% of the time, and neither 1% five bicycles to be morally wrong, and more so for the footbridge of the time, whereas in the switch dilemma they chose kill 14% of case. Critically, the moral–physical conflation hypothesis predicts the time, die 69% of the time, and neither 17% of the time; that the greater wrongness in the footbridge case will lead to a participants chose kill significantly more often in the footbridge parallel increase in participants’ choices of the causative verb kill. than switch condition, ␹2(1, N ϭ 147) ϭ 3.69, p ϭ .054, ␸ϭ.14 (see Figure 1). Number of events. There was no significant difference in Method whether participants thought multiple events (rather than a single We recruited participants using Amazon’s Mechanical Turk, event) had occurred in the footbridge (60%) and switch (61%) choosing a sample size that provides sufficient power to detect a conditions, ␹2(1, N ϭ 147) ϭ 0.00, p ϭ 1, ␸ϭ.00. large effect size. We excluded 13 participants for incomplete Moral judgment. Participants’ moral judgments showed the responses, previous participation in similar studies, or incorrectly footbridge-switch difference as expected. For ratings of moral answering a comprehension question at the end of the study (n ϭ wrongness, participants rated the actor’s behavior as more wrong ϭ ϭ ϭ 2), yielding a sample of 147 participants (Mage 32 years, 35% in the footbridge condition (M 5.05, SD 2.05) than in the female). The experimental design and stimuli were the same as in switch condition (M ϭ 3.96, SD ϭ 2.35), t(145) ϭ 3.01, p ϭ .003, Experiment 1, except that the victim was a personalized cow d ϭ 0.50. For the forced-choice question about whether the actor’s named “Lucy,” who was sacrificed to save five bicycles. We behavior was morally wrong (yes or no), participants judged the This document is copyrighted by the American Psychological Association or one of its allied publishers. personalized the cow to increase empathy and the immorality of actor’s behavior (sacrificing a personalized cow to save five bicy- This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. sacrificing the cow, motivated by recent public outrage about the cles) to be morally wrong more often in the footbridge condition killing of a lion named Cecil (McLaughlin, 2015). (76%) than in the switch condition (50%), ␹2(1, N ϭ 147) ϭ Participants were assigned to either the switch or footbridge 10.69, p ϭ .001, ␸ϭ.26. condition. In the switch condition, participants read the following: In sum, we find that when sacrificing a cow is morally question- able, participants attribute greater moral wrongness in the footbridge One day Adam was walking near some trolley tracks. Suddenly, a than switch case, and critically, they show a parallel increase in their trolley was quickly approaching out of control. Adam saw that five choice of the causative verb kill (see Figure 1). bicycles were standing on a crosswalk in the trolley’s path. Adam was standing next to a switch that could divert the trolley onto a sidetrack, but there was one cow, named Lucy, standing on the side track. Lucy Experiment 2b was dearly loved by the community and was well-known for her gentle demeanor and for raising many of the young cows in the town. In Experiment 2b, we further test whether people’s moral judg- The five bicycles were too far away and there was no other way to ments are conflated with causality by reducing the moral wrong- save them. Adam decided to pull the switch to divert the trolley. As ness of killing a personalized cow. We use the same dilemmas as 6 DE FREITAS ET AL.

in Experiment 2a but change the rescuees from bicycles back to percentages confirmed that most participants viewed sacrificing a people. We expect participants to view sacrificing the personalized personalized cow to save five people as morally permissible. cow as justified when saving five people, and if so, the moral– In sum, we found that reducing the moral wrongness of the physical conflation hypothesis predicts reduced differences be- actor’s behavior, in this case by sacrificing a personalized cow to tween footbridge and switch conditions in participants’ verb save people (rather than bicycles), also muted differences in the choices. This test also allows us to confirm that the pattern of verbs people used to describe the events. Taken together, Exper- differences found in Experiment 2a was specifically due to the iments 2a and 2b indicate that when people condemn an action, wrongness of killing a personalized cow to protect bicycles, rather they construe the causal role of the actor as direct and intended, but than the wrongness of killing a personalizing cow alone. Hence, when they want to condone an action, they choose verbs that we expect to see a similar pattern of results with sacrificing a convey indirect causality. Because, once again, the physical causal personalized cow to save five people as we saw with sacrificing a structure was the same as in Experiment 2a when bicycles (rather generic cow to save five people in Experiment 1b. than people) were at stake, these data are not adequately accounted for by the pure causality hypothesis. Instead, they support the Method moral–physical conflation hypothesis, in which changes in moral wrongness affect the verbs people choose to describe the action. We recruited participants using Amazon’s Mechanical Turk, choosing a sample size that provides sufficient power to detect a large Experiment 3 effect size. We excluded 12 participants for incomplete responses, Experiment 3 tests whether the basic footbridge-switch differ- previous participation in similar studies, or incorrectly answering a ence in verb choice also appears in participants’ spontaneous verb comprehension question at the end of the study (n ϭ 2), yielding a usage. Previous research on causality and verb choice typically sample of 148 participants (M ϭ 35 years, 56% female). The age presents participants with preset sentences which they rate or experimental design and stimuli were the same as in Experiment 2a, select (e.g., Wolff, 2003, 2007). Hence, we adopted the same except that the rescuees were people instead of bicycles. methodology in Experiments 1 and 2. These methods offer greater experimental control than examining spontaneous speech because Results and Discussion there are typically a large number of ways that a speaker can describe events. Previous research has commonly assumed that Ratings of verb accuracy. Participants’ ratings for the caus- participants’ evaluations of preset sentences will also capture pat- ative verb differed between the footbridge (M ϭ 4.84, SD ϭ 1.97) terns in spontaneous language use. and switch conditions (M ϭ 3.93, SD ϭ 2.03), t(146) ϭ 2.75, p ϭ However, it is possible that the preset sentences might not .007, d ϭ 0.45, but the difference was smaller than in Experiment 2a, reflect what participants would spontaneously produce. Hence, in and the verb accuracy ratings were lower overall. Participants’ ratings Experiment 3 we examine verb choice in a more naturalistic for the intransitive verb die did not differ between the footbridge production experiment, in which participants are asked to simply condition (M ϭ 5.92, SD ϭ 1.62) and the switch condition (M ϭ write a sentence justifying the moral judgment that they made. We 5.65, SD ϭ 1.58), t(146) ϭ 1.00, p ϭ .317, d ϭ 0.17. expected that these sentences would be more rich in content and Forced-choice of verb accuracy. In the footbridge dilemma nuanced than the preset sentences. Even so, we can observe participants chose the sentence with the causative verb kill 12% of whether participants spontaneously produce sentences that are the time, the intransitive verb die 79% of the time, and neither 8% similar to those used in Experiments 1 and 2, with a single of the time, whereas in the switch dilemma they chose kill 7% of causative verb kill or a periphrastic construction. Further, we can the time, die 79% of the time, and neither 15% of the time; the test whether verb usage differs across footbridge and switch di- proportion choosing kill did not significantly differ between con- lemmas, as in the previous experiments with preset sentences. ditions, ␹2(1, N ϭ 148) ϭ 1.38, p ϭ .239, ␸ϭ.07 (see Figure 1). Number of events. There was no significant difference in Method whether participants thought multiple events (rather than a single event) had occurred in the footbridge (60%) and switch (76%) We recruited participants using Amazon’s Mechanical Turk. We conditions, ␹2(1, N ϭ 148) ϭ 3.53, p ϭ .060, ␸ϭ.15. excluded 9 participants for incomplete responses, previous partic- This document is copyrighted by the American Psychological Association or one of its allied publishers. Moral judgment. Participants’ wrongness judgments con- ipation in similar studies, or incorrectly answering a comprehen- This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. firmed that changing the rescues from bicycles to people muted the sion question at the end of the study (n ϭ 4), yielding a sample of ϭ moral differences between footbridge and switch scenarios. For 111 participants (Mage 38 years, 58% female). ratings of moral wrongness, participants ratings did differ signif- The experimental vignettes were the same as in Experiment 1a. icantly between the footbridge condition (M ϭ 2.41, SD ϭ 1.75) Participants received either the footbridge or switch dilemma. Partic- and the switch condition (M ϭ 1.85, SD ϭ 1.45), t(146) ϭ 2.12, ipants answered whether they thought the agent was guilty of murder p ϭ .036, d ϭ 0.35, though the ratings were much lower overall (Is Adam guilty of murder? yes or no) and then explained their compared to when the cow was sacrificed to save bicycles in response (Please explain your answer in one short sentence.). These Experiment 2a. For the forced-choice question about whether the questions were designed to elicit participants’ spontaneous descrip- actor’s behavior was morally wrong (yes or no), participants’ tions of the relevant events in a natural way. Last, participants an- wrongness judgments of the actor’s behavior (sacrificing a per- swered the same comprehension and demographic items as in previ- sonalized cow to save five people) did not differ significantly ous experiments. between the footbridge (14%) and switch conditions (8%), ␹2(1, Two independent coders, who were blind to condition and the N ϭ 148) ϭ 1.25, p ϭ .264, ␸ϭ.07 (see Figure 1). These hypotheses, categorized participants’ sentences according to a rubric KILL OR DIE 7

(see the Appendix for the complete coding instructions). Participants’ inanimate subject, 0% versus 5%, ␹2(1, N ϭ 92) ϭ 2.23, p ϭ .135, responses were combined across conditions and randomly resorted for ␸ϭ.08. There was also no significant difference when we coding. The coders used the following five categories: one category grouped these three categories into a single indirect verb category, for a causative verb kill, three categories for different kinds of indirect 17% versus 16%, ␹2(1, N ϭ 92) ϭ 0.01, p ϭ .922, ␸ϭ.00. Some verbs (the intransitive verb die with and without a separate causative examples of sentences that used indirect verbs include, “He took verb, and a clause with an inanimate subject like “trolley”), and a last positive action to cause the death of another” and “He caused a category for sentences that did not fit the main categories. (Coders man to die.” Last, participants were less likely to write sentences could select more than one category if they applied to different parts in the “other” category for the footbridge (27%) than switch (80%) of a sentence but they only agreed on one such sentence.) Coders condition, ␹2(1, N ϭ 92) ϭ 25.32, p Ͻ .001, ␸ϭ.50. Evidently, received instructive feedback on the first 20 sentences from one of the sentences in the switch condition were much more likely to be authors, who was blind to condition for these sentences. Intercoder described without referring to killing or death at all. Many of these reliability for the full set of sentences was high, coders selected the sentences tended to describe the difficulty of the agent’s decision, same category for 92 of the 111 responses, giving a Cohen’s kappa emphasize that the agent did not intend to kill, and/or exonerate the value of .74. For analysis, we excluded the 19 cases in which coders agent in some way. For example, “He saved four people, made a chose different categories, yielding 92 final responses for analysis. very tough decision,” “This is a blurry line, but he did not intend The full coded dataset is available in the online supplemental material. to kill someone. He tried to save others’ lives,” and “People were going to die anyway. He reduced the fatalities.” Results and Discussion To summarize, participants spontaneously produced sentences using verbs that reflected the same distinction between direct and Forced-choice of the actor’s guilt. As expected, participants indirect causality, and these verb choices mirrored participants’ judged that the actor was guilty of murder more often in the moral judgments. Supporting the methods used in Experiments 1 footbridge condition (90%) than in the switch condition (23%), and 2, we found that nontrivial proportions of participants spon- ␹2 ϭ ϭ Ͻ ␸ϭ (1, N 92) 42.02, p .001, .65. taneously used both verb constructions that were used as stimuli in Production of direct verbs. For verb choice, we found that Experiments 1 and 2. Moreover, we observed greater use of a participants’ sentences used a lexical causative 58% of the time in single causative verb kill in the footbridge dilemma than the switch the footbridge dilemma versus 4% in the switch dilemma. Chi- dilemma, replicating this basic effect with a language production square tests showed that participants chose kill significantly more task. often in the footbridge than in the switch condition, ␹2(1, N ϭ 92) ϭ 30.22, p Ͻ .001, ␸ϭ.55 (see Figure 2). Participants wrote sentences like the following: “He willfully, knowingly murdered General Discussion him. The motive doesn’t matter” and “He still killed a man The cumulative results of Experiments 1 through 3 show that the regardless of saving five lives.” causal dynamics of a moral situation can influence the verbs Production of indirect verbs. In contrast, we did not find people use to describe it. Yet these verb choices are also sensitive significant differences across conditions for the three indirect verb to other determinants of the moral gravity of the action. Roughly categories: intransitive verb with a separate causative verb, 15% half of participants chose kill to describe the actor’s behavior in the ␹2 ϭ ϭ ϭ (footbridge) versus 11% (switch), (1, N 92) 0.21, p .647, footbridge dilemma, but after we changed the victim from a person ␸ϭ .02; intransitive verb without a separate causative, 2% versus to a cow, only 11% of participants chose kill (Experiment 1). ␹2 ϭ ϭ ϭ ␸ϭ 0%, (1, N 92) 0.93, p .336, 0.00; and clause with Experiment 2 further confirmed that participants were unlikely to use the causative verb kill if they believed the actor’s behavior was morally justifiable. In these cases, even pushing the victim off the footbridge was described as though it was an indirect cause of death, perhaps because the impact of the trolley was now construed as the more immediate cause. Yet, when participants viewed the same pushing act as immoral, they were more likely to choose the causative verb kill to match their condemnation. These effects This document is copyrighted by the American Psychological Association or one of its allied publishers. were not merely an artifact of providing preset sentences to par- This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. ticipants: Experiment 3 showed that participants’ own spontaneously written sentences showed the same asymmetry in verb choice based on moral judgment. This pattern of results points to the existence of event categories that combine elements of physical causation (relatively direct, without intervening potent links) with elements of moral responsibility (fore- seeing and intending a momentous outcome). A causative verb such as kill efficiently picks out moral transgressions from accidental, incidental, or inconsequential events, and communicating this choice of event category can be used rhetorically and tendentiously to influ- Figure 2. Participants’ judgments about whether the agent was guilty of ence how others behave (Pinker, 2007). When expressing a public murder and their usage of a single causative verb for kill to describe the protest, weaving a narrative before a jury, crafting a newspaper agent’s action in Experiment 3. headline, or gossiping about everyday offenses, people can choose a 8 DE FREITAS ET AL.

lexical causative verb that implicates intentional, direct causality to form, which may be contrasted for rhetorical or moral effect, for bolster their moral accusation, as in That surgeon killed my husband example, Mike injured him versus Mike caused/allowed him to be or the famous 1960s Vietnam protest chant, “Hey, hey, LBJ, how injured and Sally broke the bicycle versus Sally caused the bicycle to many kids did you kill today?” Conversely, they will opt for peri- break. More generally, we believe that the intersection between moral phrastic constructions containing generic causative verbs such as psychology and psycholinguistics can explain many word choices in make, lead to,orresult in when they seek to conceal agency and language production, and holds the promise of illuminating issues in hence responsibility, as in the notorious euphemism caused collateral the study of composition, rhetoric, persuasion, and style. damage and the evasive mistakes were made. We conclude that the potency of verbs in such incidences and in moral reasoning more generally is due to the existence of cognitive categories which are References essential in social discourse by singling out events for which speakers Alicke, M. D. (1992). Culpable causation. Journal of Personality and feel they have a right to hold the agents morally responsible, namely Social Psychology, 63, 368–378. http://dx.doi.org/10.1037/0022-3514 events whose outcomes are foreseen, intended, and directly caused .63.3.368 without an intervening causal force. Berinsky, A. J., Huber, G. A., & Lenz, G. S. (2012). Evaluating online labor This influence of morality on verb choice is predicted by the markets for experimental research: Amazon.com’s Mechanical turk. Political moral–physical conflation hypothesis. The pure causality hypoth- Analysis, 20, 351–368. http://dx.doi.org/10.1093/pan/mpr057 esis, in contrast, emphasizes only the physics of the interaction, Buhrmester, M., Kwang, T., & Gosling, S. D. (2011). Amazon’s Mechan- and so cannot account for the large effect of moral judgment on ical Turk: A new source of inexpensive, yet high-quality data? Perspec- verb choice. This is not to say that physics does not also play an tives on Psychological Science, 6, 3–5. http://dx.doi.org/10.1177/ 1745691610393980 important role. Previous research already showed that physical De Freitas, J., & Johnson, S. G. B. (2015). Behaviorist thinking in judgments of causality affects verb choice (e.g., Wolff, 2003, 2007). Indeed, the wrongness, punishment, and blame. In D. C. Noelle, R. Dale, A. S. War- fact that participants chose verbs that portray a more direct and laumont, J. Yoshimi, T. Matlock, C. D. Jennings, & P. P. Maglio (Eds.), hence culpable physical role for the agent implies that they expect In Proceedings of the 37th annual conference of the Cognitive Science listeners’ moral judgments to be swayed by physical causality. The Society (pp. 524–529). Austin, TX: Cognitive Science Society. novelty of the current contribution can thus be summarized in two DeScioli, P., Bruening, R., & Kurzban, R. (2011). The omission effect in mutually reinforcing discoveries: (a) moral judgments influence moral cognition: Toward a functional explanation. Evolution and Human verb choice, and (b) verb choice, in turn, can reveal how moral Behavior, 32, 204–215. http://dx.doi.org/10.1016/j.evolhumbehav.2011 judgments interact with causal cognition. .01.003 One interesting question about the present results is why par- DeScioli, P., & Kurzban, R. (2009). Mysteries of morality. Cognition, 112, ticipants tended to judge the footbridge actor as wrong more often, 281–299. http://dx.doi.org/10.1016/j.cognition.2009.05.008 DeScioli, P., & Kurzban, R. (2013). A solution to the mysteries of morality. overall, than they chose kill over cause to die. That is, why did Psychological Bulletin, 139, 477–496. http://dx.doi.org/10.1037/ some participants condemn the agent but still show some resis- a0029065 tance to saying he killed the victim? This reluctance might reflect Foot, P. (1978). Virtues and vices and other essays in moral philosophy. how people grapple with the dilemma. Although they judge that Berkeley, CA: University of California Press. killing a man is wrong, they also understand that the intention Gergely, G., & Bever, T. G. (1986). Related intuitions and the mental behind it was to save five lives. The forced-choice wrongness representation of causative verbs in adults and children. Cognition, 23, measure did not allow participants to express these mixed feelings, 211–277. whereas some who judged the action as wrong might have still felt Goldberg, A. E. (2001). Patient arguments of causative verbs can be that the cause to die sentence better captured their mixture of omitted: The role of information structure in argument distribution. judgment and understanding of good intentions. Furthermore, be- Language Sciences, 23, 503–524. cause cause to die can be an acceptable description for both direct Goodman, J. K., Cryder, C. E., & Cheema, A. (2012). Data collection in a and indirect causation (Wolff, 2003), participants did not need to flat world: The strengths and weaknesses of Mechanical Turk samples. Journal of Behavioral Decision Making, 26, 213–224. http://dx.doi.org/ choose kill unless they wanted to emphasize direct causation. 10.1002/bdm.1753 One question for future work is how people choose verbs in the Greene, J. (2014). Moral tribes: Emotion, reason and the gap between us midst of communication, debate, and negotiation about moral wrong-

This document is copyrighted by the American Psychological Association or one of its allied publishers. and them. New York, NY: Penguin. doing. The present experiments used simple, controlled tasks in which Greene, J. D., Sommerville, R. B., Nystrom, L. E., Darley, J. M., & Cohen, This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. participants evaluated sentences or explained their judgments in a J. D. (2001). An fMRI investigation of emotional engagement in moral single statement. Even in this minimal social context, moral judgment judgment. Science, 293, 2105–2108. http://dx.doi.org/10.1126/science influenced verb choice. However, the interactions between verbs, .1062872 causality, and moral judgment are likely to be even more complex in Haidt, J. (2001). The emotional dog and its rational tail: A social intuition- back-and-forth dialogue about an agent’s behavior. For instance, a ist approach to moral judgment. Psychological Review, 108, 814–834. speaker might exaggerate direct verbs to threaten condemnation or to http://dx.doi.org/10.1037/0033-295X.108.4.814 recruit others against the accused. Alternatively, a speaker could offer Haspelmath, M. (1993). More on the typology of inchoative/causative verb alternations. In B. Comrie & M. Polinsky (Eds.), Causatives and tran- indirect verbs to convey a willingness to compromise and assign only sitivity. Amsterdam, the Netherlands: John Benjamins. http://dx.doi.org/ moderate responsibility to the agent. Finally, future research can 10.1075/slcs.23.05has examine other morally relevant verbs in addition to kill and die. Hauser, M., Cushman, F., Young, L., Kang-Xing Jin, R., & Mikhail, J. Although there are not many mutually exclusive verb pairs like kill (2007). A dissociation between moral judgments and justifications. Mind and die, there are many examples of lexical versus periphrastic caus- & Language, 22, 1–21. http://dx.doi.org/10.1111/j.1468-0017.2006 atives with the same verbs, sometimes in the passive or intransitive .00297.x KILL OR DIE 9

Ipeirotis, P. (2010). The new demographics of Mechanical Turk. Retrieved Mikhail, J. (2007). Universal moral grammar: Theory, evidence and the from http://www.behind-the-enemy-lines.com/2010/03/new-demo future. Trends in Cognitive Sciences, 11, 143–152. http://dx.doi.org/10 graphics-of-mechanical-turk.html .1016/j.tics.2006.12.007 Knobe, J. (2005). Theory of mind and moral cognition: Exploring the Paolacci, G., Chandler, J., & Ipeirotis, P. G. (2010). Running experiments on connections. Trends in Cognitive Sciences, 9, 357–359. http://dx.doi.org/ Amazon Mechanical Turk. Judgment and Decision Making, 5, 411–419. 10.1016/j.tics.2005.06.011 Pinker, S. (1989). Learnability and cognition: The acquisition of argument Kurzban, R., DeScioli, P., & Fein, D. (2012). Hamilton vs. Kant: Pitting structure. Cambridge, MA: MIT Press. adaptations for altruism against adaptations for moral judgment. Evolu- Pinker, S. (2007). The stuff of thought: Language as a window into human tion and Human Behavior, 33, 323–333. http://dx.doi.org/10.1016/j nature. Penguin. .evolhumbehav.2011.11.002 Spellman, B. A., & Mandel, D. R. (1999). When possibility informs reality: Levin, B. (1993) English verb classes and alternations: A preliminary Counterfactual thinking as a cue to causality. Current Directions in investigation. Chicago, IL: University of Chicago Press. Psychological Science, 8, 120–123. Malle, B. F. (2006). Intentionality, morality, and their relationship in Wolff, P. (2003). Direct causation in the linguistic coding and individua- human judgment. Journal of Cognition and Culture, 6, 87–112. http:// tion of causal events. Cognition, 88, 1–48. http://dx.doi.org/10.1016/ dx.doi.org/10.1163/156853706776931358 S0010-0277(03)00004-0 McLaughlin, E. C. (2015, October 12). Zimbabwe won’t press charges Wolff, P. (2007). Representing causation. Journal of Experimental Psy- against Cecil the Lion’s killer. Retrieved from http://www.cnn.com/ chology: General, 136, 82–111. http://dx.doi.org/10.1037/0096-3445 2015/10/12/africa/zimbabwe-cecil-lion-walter-palmer-no-charges/ .136.1.82

Appendix Instructions Given to Coders in Experiment 3

The following instructions were provided independently to two Grammar Primer hypothesis-blind coders, who used the instructions to code the sentences produced by participants in Experiment 3. Here are a few basic grammatical distinctions to help you code In this task, you will read responses written by study participants the grammar of participants’ responses. ϩ and code the grammar that participants used in their response. • Transitive verb. A transitive verb takes the form subject verb ϩ object, such as Alice kicked the ball,orAlice killed The Study the spider. • Intransitive verb. An intransitive verb takes the form Participants read some scenarios about Adam, who faced a subject ϩ verb, without requiring an object, such as Alice dilemma involving five people, one man, and an oncoming trolley. jumped or The spider died. If Adam acted one way, five people would die; if Adam acted • Passive voice. A verb in the passive voice appears in the another way, one man would die. Adam made the choice, and one participle form (usually with an —en or –ed suffix), and man died. (We are leaving out the specifics of the scenarios so they has the done-to or acted-upon as the subject, while either do not influence coding.) omitting the doer or actor or expressing it in a by-phrase as Participants then answered the questions: in The ball was kicked, The cake was eaten by the boy, or The woman was hit by the baby. 1. Is Adam guilty of murder? (yes or no) • Causal verb. A causal verb is about causality, including 2. Please explain your answer in one short sentence. caused/let/made/allowed, and is often followed by another verb that is intransitive or in the passive voice. For exam- You will read participants’ responses to Question 2 and code the ple, Alice made the ball roll, Alice caused the spider to die, grammar that they used to refer to the man’s death. or Alice allowed the spider to be killed. This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

(Appendix continues) 10 DE FREITAS ET AL.

Coding Instructions Examples of acceptable variations: Adam/he caused/let/made/ allowed the man to die/be killed/get run over. Adam’s/his action Read each participant’s sentence(s). Some responses might be led to the death of the man. long and wordy, but focus on the part(s) of the sentence that refers (3) Adam . . . and the man died. to the man’s death. We want you to code the grammatical structure This category is the same as (2), except that the clause does not of that part by using the rubric below. involve a causative verb like caused/let/made/allowed/led. Record whether that part most closely fits one of the categories Examples of acceptable variations: Adam/he . . . and the man below by putting a “1” in the categories’ column within the excel died/was killed/was run over as a result/as a consequence. sheet. Each response will typically fit best in exactly one category, (4) the trolley killed the man. except in rare cases in which two different parts of the sentence fit The sentence includes a clause with the trolley or another different categories; in these rare cases, you can select multiple inanimate thing as the subject rather than Adam, and the man as categories. the direct object of the verb in that clause. That clause can also be a subordinate clause in a sentence in which the main clause refers to what Adam did. Coding Rubric Examples of acceptable variations: Adam/he . . . caused the (1) Adam killed the man. trolley to kill the man; Adam/he . . . and the trolley killed the man. The sentence includes a single clause with the subject Adam, a (5) Other transitive verb meaning kill, and the man as the direct object of the Pick this category if the sentence does not fall into any of the first four categories. The sentence does not refer to the man’s verb. death. It might restate that Adam made a tough choice, without Examples of acceptable variations: Adam/He killed/murdered/ specifically saying the man died. Or it might say that Adam pushed sacrificed the man. Adam/He Committed murder/manslaughter. the man or flipped a switch, also without saying the man died. Also (2) Adam caused the man to die. included in this category are idioms, such as “pushed the man to The sentence includes a main clause with the subject Adam, a his death” and “took the man’s life.” verb about causality, the man, and a subordinate clause with a separate verb (either an intransitive verb in the active voice, such Received February 18, 2016 as die, or a transitive verb in the passive voice, such as be killed or Revision received October 18, 2016 get run over), for the man’s death. Accepted October 21, 2016 Ⅲ This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.